Tag
This paper introduces DiaLLM, a framework for adapting LLMs to English dialects, revealing a gap between dialectal robustness (understanding) and generation (producing dialectal text), and showing that explicit variety-targeted alignment improves generation but not necessarily human preference.